On the evaluation of defect prediction models